Caveman: Make Your AI Coding Agent Say Less and Read Less
Shorter answers, smaller inputs, same code.
The caveman voice trims chatty answers while keeping code, commands, paths and error messages exact. Input compression shrinks logs, JSON, CSV and test output before the model reads them, and keeps the originals on disk.
The Six Parts of Caveman 4.0
All are on after install; each has its own switch:
The caveman voice: answer first, one idea per sentence, payload kept verbatim.
Compresses logs, JSON, CSV and test output before the model reads them.
Picks the model and effort for each request. Needs a free account.
Reusable scripts for code mode, so agents stop retyping the same helper (pre-release).
Finds where tokens go and keeps the prompt cache warm while you pause.
Returns web pages to the agent in compressed form.
Outside Results Cited in the README
| Who | Result |
|---|---|
| Adobe Research (CAVEWOMAN paper) | Caveman-style output cut cost 1.4× to 2.4× per model, up to 3×, across eight models and five datasets |
| Elastic (Elasticsearch Labs) | 63.6% fewer response tokens in eight live MCP scenarios |
| JetBrains | No measurable quality loss (p = 0.82) and 8.5% fewer output tokens on 86 coding tasks |
Input Compression by File Type (Project Measurements)
| File type | Whole-session saving |
|---|---|
| CSV | 55.1% |
| Logs | 50.2% |
| YAML | 46.2% |
| Test output | 27.8% |
| JSON | 26.4% |
| HTML | 9.9% worse (not compressed yet) |
| All six together | 33.2% fewer input tokens |
One-Command Install
npx @caveman-ai/cli@latest
Mac, Linux and Windows; needs Node.js 22.13+. It finds your agents and logins, shows what it will change, and asks you to confirm. Afterwards plain claude runs with caveman on. caveman status shows what is enabled.
npx skills add JuliusBrussee/caveman -g
claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman
caveman off browse
npx -y github:JuliusBrussee/caveman -- --uninstall
Why Your Prompts Are Left Alone
The README cites Adobe Research finding that compressing the user's own prompt makes answers worse. So caveman never rewrites what you type; it shortens what the agent says and what it reads from tools.
When It Saves Less
The project's own table shows HTML input getting slightly worse, and notes that "Be concise" is already a strong baseline on new models. Measure on your own sessions with caveman status and your provider's usage data.
Honest Numbers
Caveats the README itself states:
- Interim result: The Auto "54% cheaper" figure comes from a half-finished run on 85 tasks.
- Older measurements: The session compression run dates from August 2026 on an older engine.
- Pre-release parts: Blocks is a release candidate, Python only, with no sandbox yet.
- Account for Auto: Only Auto needs an account; without signing in, it is off and nothing is sent for it.
Uses your own keys or Claude Pro/Max login on your machine. Originals of compressed files stay on disk.
Apache-2.0 (per GitHub repository metadata).
README and INSTALL.md cover 30+ agents, Windows, uninstall and the release notes.
Frequently Asked Questions
Does Caveman change the code my agent writes?
No. The README says code, commands, paths and error messages stay untouched, character for character. Only the conversational text gets shorter.
How do I install Caveman?
Run npx @caveman-ai/cli@latest (Node.js 22.13 or newer). It finds your agents, shows what it will change, and asks you to confirm. If you only want shorter answers, run npx skills add JuliusBrussee/caveman -g instead.
Does Caveman rewrite my prompts?
No. The project says your prompts are never rewritten, because research it cites found that compressing the user's prompt makes answers worse.
Does Caveman send my data anywhere?
It runs on your machine with your own keys or login. The optional Auto model router needs a free account; if you are not signed in, Auto is off and nothing is sent for it. Read the privacy section of the README for details.
